HUGO [1] and MOD [2] are the most secure adaptive embedding algorithms of 2011. These algorithms strive to hide a secret message, while minimizing an ad hoc embedding impact. They use a detectability map, which, if properly defined, is correlated to the security. In this paper, we present a new adaptive embedding scheme: Adaptive Steganography by Oracle (ASO). It is based on an oracle used to calculate the detectability map, and this oracle use the Kodovsky's ensemble classifiers [3]. Our approach preserves both cover image and sender's database distributions during the embedding process, which improves the security. In addition, it offers to the sender the opportunity to choose the most reliable image(s), during his secret communication. Experimental results show that our embedding scheme presents good secu-rity performances, since the detection error of ASO is much higher than that of HUGO.